Working in the field of data science, I’m always seeking ways to keep current in the field and there are a number of important resources available for this purpose: new book titles, blog articles, conference sessions, Meetups, webinars/podcasts, not to mention the gems floating around in social media. But to dig even deeper, I routinely look at what’s coming out of the world’s research labs. And one great way to keep a pulse for what the research community is working on is to monitor the flow of new machine learning Ph.D. dissertations. Admittedly, many such theses are laser-focused and narrow, but from previous experience reading these documents, you can learn an awful lot about new ways to solve difficult problems over a vast range of problem domains.
In this article, I present a number of hand-picked machine learning dissertations that I found compelling in terms of my own areas of interest and aligned with problems that I’m working on. I hope you’ll find a number of them that match your own interests. Each dissertation may be challenging to consume but the process will result in hours of satisfying summer reading. Enjoy!
Please check out my previous data science dissertation round-up article.
This machine learning dissertation concerns the geometric problem of finding a convex set that best fits a given data set. The overarching question serves as an abstraction for data-analytical tasks arising in a range of scientific and engineering applications with a focus on two specific instances: (i) a key challenge that arises in solving inverse problems is ill-posedness due to a lack of measurements. A prominent family of methods for addressing such issues is based on augmenting optimization-based approaches with a convex penalty function so as to induce a desired structure in the solution. These functions are typically chosen using prior knowledge about the data. The thesis also studies the problem of learning convex penalty functions directly from data for settings in which we lack the domain expertise to choose a penalty function. The solution relies on suitably transforming the problem of learning a penalty function into a fitting task; and (ii) the problem of fitting tractably-described convex sets given the optimal value of linear functionals evaluated in different directions.
This machine learning dissertation analyzes data to build a quantitative understanding of the world. Linear algebra is the foundation of algorithms, dating back one hundred years, for extracting structure from data. Modern technologies provide an abundance of multi-dimensional data, in which multiple variables or factors can be compared simultaneously. To organize and analyze such data sets we can use a tensor, the higher-order analogue of a matrix. However, many theoretical and practical challenges arise in extending linear algebra to the setting of tensors. The first part of the thesis studies and develops the algebraic theory of tensors. The second part of the thesis presents three algorithms for tensor data. The algorithms use algebraic and geometric structure to give guarantees of optimality.
This machine learning dissertation is primarily a description of three projects. It starts with a method for spatial prediction and parameter estimation for irregularly spaced, and non-Gaussian data. It is shown that by judiciously replacing the likelihood with an empirical likelihood in the Bayesian hierarchical model, approximate posterior distributions for the mean and covariance parameters can be obtained. Due to the complex nature of the hierarchical model, standard Markov chain Monte Carlo methods cannot be applied to sample from the posterior distributions. To overcome this issue, a generalized sequential Monte Carlo algorithm is used. Finally, this method is applied to iron concentrations in California. The second project focuses on anomaly detection for functional data; specifically for functional data where the observed functions may lie over different domains. By approximating each function as a low-rank sum of spline basis functions the coefficients will be compared for each basis across each function. The idea being, if two functions are similar then their respective coefficients should not be significantly different. This project concludes with an application of the proposed method to detect anomalous behavior of users of a supercomputer at NREL. The final project is an extension of the second project to two-dimensional data. This project aims to detect location and temporal anomalies from ground motion data from a fiber-optic cable using distributed acoustic sensing (DAS).
Advances in data acquisition technology pose challenges in analyzing large volumes of streaming data. Sampling is a natural yet powerful tool for analyzing such data sets due to their competent estimation accuracy and low computational cost. Unfortunately, sampling methods and their statistical properties for streaming data, especially streaming time series data, are not well studied in the literature. Meanwhile, estimating the dependence structure of multidimensional streaming time-series data in real-time is challenging. With large volumes of streaming data, the problem becomes more difficult when the multidimensional data are collected asynchronously across distributed nodes, which motivates us to sample representative data points from streams. This machine learning dissertation proposes a series of leverage score-based sampling methods for streaming time series data. The simulation studies and real data analysis are conducted to validate the proposed methods. The theoretical analysis of the asymptotic behaviors of the least-squares estimator is developed based on the subsamples.
This machine learning dissertation develops statistical machine learning methodology for three distinct tasks. Each method blends classical statistical approaches with machine learning methods to provide principled solutions to problems with complex, heterogeneous data sets. The first framework proposes two methods for high-dimensional shape-constrained regression and classification. These methods reshape pre-trained prediction rules to satisfy shape constraints like monotonicity and convexity. The second method provides a nonparametric approach to the econometric analysis of discrete choice. This method provides a scalable algorithm for estimating utility functions with random forests, and combines this with random effects to properly model preference heterogeneity. The final method draws inspiration from early work in statistical machine translation to construct embeddings for variable-length objects like mathematical equations
This machine learning dissertation comprises four chapters. The first is an introduction to the topics of the dissertation and the remaining chapters contain the main results. Chapter 2 gives new results for consistency of maximum likelihood estimators with a focus on multivariate mixed models. The presented theory builds on the idea of using subsets of the full data to establish consistency of estimators based on the full data. The theory is applied to two multivariate mixed models for which it was unknown whether maximum likelihood estimators are consistent. In Chapter 3 an algorithm is proposed for maximum likelihood estimation of a covariance matrix when the corresponding correlation matrix can be written as the Kronecker product of two lower-dimensional correlation matrices. The proposed method is fully likelihood-based. Some desirable properties of separable correlation in comparison to separable covariance are also discussed. Chapter 4 is concerned with Bayesian vector auto-regressions (VARs). A collapsed Gibbs sampler is proposed for Bayesian VARs with predictors and the convergence properties of the algorithm are studied.
In the era of big data, uncovering useful information and hidden patterns in the data is prevalent in different fields. However, it is challenging to effectively select input variables in data and estimate their effects. The goal of this machine learning dissertation is to develop reproducible statistical approaches that provide mechanistic explanations of the phenomenon observed in big data analysis. The research contains two parts: variable selection and model estimation. The first part investigates how to measure and interpret the usefulness of an input variable using an approach called “variable importance learning” and builds tools (methodology and software) that can be widely applied. Two variable importance measures are proposed, a parametric measure SOIL and a non-parametric measure CVIL, using the idea of a model combining and cross-validation respectively. The SOIL method is theoretically shown to have the inclusion/exclusion property: When the model weights are properly around the true model, the SOIL importance can well separate the variables in the true model from the rest. The CVIL method possesses desirable theoretical properties and enhances the interpretability of many mysterious but effective machine learning methods. The second part focuses on how to estimate the effect of a useful input variable in the case where the interaction of two input variables exists. Investigated is the minimax rate of convergence for regression estimation in high-dimensional sparse linear models with two-way interactions, and construct an adaptive estimator that achieves the minimax rate of convergence regardless of the true heredity condition and the sparsity indices.
While the term “Big Data” can have multiple meanings, this dissertation considers the type of data in which the number of features can be much greater than the number of observations (also known as high-dimensional data). High-dimensional data is abundant in contemporary scientific research due to the rapid advances in new data-measurement technologies and computing power. Recent advances in statistics have witnessed great development in the field of high-dimensional data analysis. This machine learning dissertation proposes three methods that study three different components of a general framework of the high-dimensional structured regression problem. A general theme of the proposed methods is that they cast a certain structured regression as a convex optimization problem. In so doing, the theoretical properties of each method can be well studied, and efficient computation is facilitated. Each method is accompanied by a thorough theoretical analysis of its performance, and also by an R package containing its practical implementation. It is shown that the proposed methods perform favorably (both theoretically and practically) compared with pre-existing methods.
Decision trees and decision tree ensembles are widely used nonparametric statistical models. A decision tree is a binary tree that recursively segments the covariate space along the coordinate directions to create hyper rectangles as basic prediction units for fitting constant values within each of them. A decision tree ensemble combines multiple decision trees, either in parallel or in sequence, in order to increase model flexibility and accuracy, as well as to reduce prediction variance. Despite the fact that tree models have been extensively used in practice, results on their asymptotic behaviors are scarce. This machine learning dissertation presents analyses on tree asymptotics in the perspectives of tree terminal nodes, tree ensembles, and models incorporating tree ensembles respectively. The study introduces a few new tree-related learning frameworks which provides provable statistical guarantees and interpretations. A study on the Gini index used in the greedy tree building algorithm reveals its limiting distribution, leading to the development of a test of better splitting that helps to measure the uncertain optimality of a decision tree split. This test is combined with the concept of decision tree distillation, which implements a decision tree to mimic the behavior of a block box model, to generate stable interpretations by guaranteeing a unique distillation tree structure as long as there are sufficiently many random sample points. Also applied is mild modification and regularization to the standard tree boosting to create a new boosting framework named Boulevard. Also included is an integration of two new mechanisms: honest trees, which isolate the tree terminal values from the tree structure, and adaptive shrinkage, which scales the boosting history to create an equally weighted ensemble. This theoretical development provides the prerequisite for the practice of statistical inference with boosted trees. Lastly, the thesis investigates the feasibility of incorporating existing semi-parametric models with tree boosting.
This dissertation develops Bayesian methods for handling unit nonresponse, item nonresponse, and erroneous responses in large-scale surveys and censuses containing categorical data. The focus is on applications of nested household data where individuals are nested within households and certain combinations of the variables are not allowed, such as the U.S. Decennial Census, as well as surveys subject to both unit and item nonresponse, such as the Current Population Survey.
Due to recent advances in computer technology, the cost of collecting and storing data has dropped drastically. This makes it feasible to collect large amounts of information for each data point. This increasing trend in feature dimensionality justifies the need for research on variable selection. Random forest (RF) has demonstrated the ability to select important variables and model complex data. However, simulations confirm that it fails in detecting less influential features in presence of variables with large impacts in some cases. This dissertation proposes two algorithms for localized variable selection: clustering-based feature selection (CBFS) and locally adjusted feature importance (LAFI). Both methods aim to find regions where the effects of weaker features can be isolated and measured. CBFS combines RF variable selection with a two-stage clustering method to detect variables where their effect can be detected only in certain regions. LAFI, on the other hand, uses a binary tree approach to split data into bins based on response variable rankings, and implements RF to find important variables in each bin. Larger LAFI is assigned to variables that get selected in more bins. Simulations and real data sets are used to evaluate these variable selection methods.
The focus of this dissertation is on functional data which are sparsely and irregularly observed. Such data require special consideration, as classical functional data methods and theory were developed for densely observed data. As is the case in much of functional data analysis, the functional principal components (FPCs) play a key role in current sparse functional data methods via the Karhunen-Loéve expansion. Thus, after a review of relevant background material, this dissertation is divided roughly into two parts, the first focusing specifically on theoretical properties of FPCs, and the second on regression for sparsely observed functional data.
In observational studies, identifying assumptions may fail, often quietly and without notice, leading to biased causal estimates. Although less of a concern in randomized trials where treatment is assigned at random, bias may still enter the equation through other means. This dissertation has three parts, each developing new methods to address a particular pattern or source of bias in the setting being studied. The first part extends the conventional sensitivity analysis methods for observational studies to better address patterns of heterogeneous confounding in matched-pair designs. The second part develops a modified difference-in-difference design for comparative interrupted time-series studies. The method permits partial identification of causal effects when the parallel trends assumption is violated by an interaction between group and history. The method is applied to a study of the repeal of Missouri’s permit-to-purchase handgun law and its effect on firearm homicide rates. The final part presents a study design to identify vaccine efficacy in randomized control trials when there is no gold standard case definition. The approach augments a two-arm randomized trial with natural variation of a genetic trait to produce a factorial experiment.
Sparsity is a standard structural assumption that is made while modeling high-dimensional statistical parameters. This assumption essentially entails a lower-dimensional embedding of the high-dimensional parameter thus enabling sound statistical inference. Apart from this obvious statistical motivation, in many modern applications of statistics such as Genomics, Neuroscience, etc. parameters of interest are indeed of this nature. For over almost two decades, spike and slab type priors have been the Bayesian gold standard for modeling of sparsity. However, due to their computational bottlenecks, shrinkage priors have emerged as a powerful alternative. This family of priors can almost exclusively be represented as a scale mixture of Gaussian distribution and posterior Markov chain Monte Carlo (MCMC) updates of related parameters are then relatively easy to design. Although shrinkage priors were tipped as having computational scalability in high-dimensions, when the number of parameters is in thousands or more, they do come with their own computational challenges. Standard MCMC algorithms implementing shrinkage priors generally scale cubic in the dimension of the parameter making real-life application of these priors severely limited.
The first chapter of this dissertation addresses this computational issue and proposes an alternative exact posterior sampling algorithm complexity of which that linearly in the ambient dimension. The algorithm developed in the first chapter is specifically designed for regression problems. The second chapter develops a Bayesian method based on shrinkage priors for high-dimensional multiple response regression. Chapter three chooses a specific member of the shrinkage family known as the horseshoe prior and studies its convergence rates in several high-dimensional models.
This dissertation proposes novel methods to tackle two problems: the misspecified model with measurement error and high-dimensional binary classification, both have a crucial impact on applications in public health. The first problem exists in the epidemiology practice. Epidemiologists often categorize a continuous risk predictor since categorization is thought to be more robust and interpretable, even when the true risk model is not a categorical one. Thus, their goal is to fit the categorical model and interpret the categorical parameters. The second project considers the problem of high-dimensional classification between the two groups with unequal covariance matrices. Rather than estimating the full quadratic discriminant rule, it is proposed to perform simultaneous variable selection and linear dimension reduction on original data, with the subsequent application of quadratic discriminant analysis on the reduced space. Further, in order to support the proposed methodology, two R packages were developed, CCP and DAP, along with two vignettes as long-format illustrations for their usage.
This dissertation contains three chapters that consider penalized regression from a model-based perspective, interpreting penalties as assumed prior distributions for unknown regression coefficients. The first chapter shows that treating a lasso penalty as a prior can facilitate the choice of tuning parameters when standard methods for choosing the tuning parameters are not available, and when it is necessary to choose multiple tuning parameters simultaneously. The second chapter considers a possible drawback of treating penalties as models, specifically possible misspecification. The third chapter introduces structured shrinkage priors for dependent regression coefficients which generalize popular independent shrinkage priors. These can be useful in various applied settings where many regression coefficients are not only expected to be nearly or exactly equal to zero, but also structured.
This dissertation revisits and makes progress on some old but challenging problems concerning least squares estimation, the work-horse of supervised machine learning. Two major problems are addressed: (i) least squares estimation with heavy-tailed errors, and (ii) least squares estimation in non-Donsker classes. For (i), this problem is studied both from a worst-case perspective, and a more refined envelope perspective. For (ii), two case studies are performed in the context of (a) estimation involving sets and (b) estimation of multivariate isotonic functions. Understanding these particular aspects of least squares estimation problems requires several new tools in the empirical process theory, including a sharp multiplier inequality controlling the size of the multiplier empirical process, and matching upper and lower bounds for empirical processes indexed by non-Donsker classes.
How to Learn More about Machine Learning
At our upcoming event this November 16th-18th in San Francisco, ODSC West 2021 will feature a plethora of talks, workshops, and training sessions on machine learning and machine learning research. You can register now for 50% off all ticket types before the discount drops to 40% in a few weeks. Some highlighted sessions on machine learning include:
- Towards More Energy-Efficient Neural Networks? Use Your Brain!: Olaf de Leeuw | Data Scientist | Dataworkz
- Practical MLOps: Automation Journey: Evgenii Vinogradov, PhD | Head of DHW Development | YooMoney
- Applications of Modern Survival Modeling with Python: Brian Kent, PhD | Data Scientist | Founder The Crosstab Kite
- Using Change Detection Algorithms for Detecting Anomalous Behavior in Large Systems: Veena Mendiratta, PhD | Adjunct Faculty, Network Reliability and Analytics Researcher | Northwestern University
Sessions on MLOps:
- Tuning Hyperparameters with Reproducible Experiments: Milecia McGregor | Senior Software Engineer | Iterative
- MLOps… From Model to Production: Filipa Peleja, PhD | Lead Data Scientist | Levi Strauss & Co
- Operationalization of Models Developed and Deployed in Heterogeneous Platforms: Sourav Mazumder | Data Scientist, Thought Leader, AI & ML Operationalization Leader | IBM
- Develop and Deploy a Machine Learning Pipeline in 45 Minutes with Ploomber: Eduardo Blancas | Data Scientist | Fidelity Investments
Sessions on Deep Learning:
- GANs: Theory and Practice, Image Synthesis With GANs Using TensorFlow: Ajay Baranwal | Center Director | Center for Deep Learning in Electronic Manufacturing, Inc
- Machine Learning With Graphs: Going Beyond Tabular Data: Dr. Clair J. Sullivan | Data Science Advocate | Neo4j
- Deep Dive into Reinforcement Learning with PPO using TF-Agents & TensorFlow 2.0: Oliver Zeigermann | Software Developer | embarc Software Consulting GmbH
- Get Started with Time-Series Forecasting using the Google Cloud AI Platform: Karl Weinmeister | Developer Relations Engineering Manager | Google